Dataset for "Machine-learning-accelerated screening of experimentally reported covalent organic frameworks for selective CO₂ capture"

- Input structures: the 874 frameworks of the CURATED-COFs database (Ongari et al., ACS Cent. Sci. 2019, 5, 1663; ACS Cent. Sci. 2020, 6, 1890). - Charges: EQeq framework charges were computed for 871 frameworks. The 3 frameworks with more than 3,000 atoms per cell were excluded. - Adsorption: CO2 and N2 Henry coefficients, Henry selectivities and zero-coverage isosteric heats at 298 K were computed by Widom insertion (RASPA 2; DREIDING/UFF framework; TraPPE gases) for 866 frameworks. - Screening set: after quality control, 839 frameworks remain. Excluded were 10 with unphysical EQeq charges, 5 containing Li+/Na+ counter-ions, and 12 with no CO2-accessible surface. - Validation: pure-CO2 GCMC runs for 9 frameworks with experimental data. - DFT: counterpoise-corrected PBE-D3(BJ)/def2-TZVP interaction energies of CO2 and N2 with 5 linkage-motif clusters and a benzene reference. - Machine learning: random forest, MLP and SchNet predictions (5-fold cross-validation and a 100-framework screening scenario).